Learning 4D Geometric Priors for Inference-Efficient World Action Models
1. Paper overview
In one sentence: MECo-WAM uses temporary geometric guidance and action-weighted relational supervision during training to improve manipulation while retaining the base video-action inference graph. e-probleme-architecturee-attentione-distillatione-robotwine-depthe-ablation
| At a glance | What to know |
|---|---|
| Research problem | Author claim Appearance-oriented video prediction may miss the object–gripper–target relations needed for precise manipulation. The paper asks whether evolving geometry can improve policy representations during training while preserving a lightweight observation-to-action deployment graph. e-problem |
| Core mechanism | Source description A removable 4D expert obtains geometric targets from frozen VGGT, alongside video and action denoising. e-architecture |
| A key reported result | RoboTwin 2.0 manipulation: MECo-WAM: clean 93.26, randomized 91.98, average 92.62. Task success rate (%). Clean/randomized bimanual simulation following the cited Fast-WAM configuration; both compared methods lack embodied-policy pretraining. Fast-WAM: 91.88, 91.78, 91.83; average gain 0.79 percentage points. Pretrained LingBot-VA averages 92.20. Clean gain: 1.38 points; randomized gain: 0.20. No uncertainty is reported. e-protocole-robotwin |
| Reading caution | Reader analysis Small simulation gains lack reported seeds, confidence intervals and rollout counts. Real-world evaluation covers two cube tasks; Table 3 shows Test 1, Test 2 and Test n without defining n, missing-value handling or all metric denominators. e-protocole-robotwine-reale-ablation |
Core contributions
- Source description
A removable 4D expert obtains geometric targets from frozen VGGT, alongside video and action denoising. e-architecture
- Author claim
Stochastically decayed access to current geometry and action-aware spatial/temporal losses are proposed to transfer geometry without future-information shortcuts. e-attentione-distillation
Figure 2. Geometry is an additional training pathway that is removed before action inference. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the left panel from the input boxes upward. Current and future RGB frames supply the video stream; the action chunk supplies the action stream; frozen VGGT encodes frames for geometry supervision. The current anchors remain clean, while dashed noise boxes surround future slots. Predicted video, action and geometry outputs appear above their experts. Cross the vertical divider to see what deployment retains: observation, instruction and robot state condition an action chunk through video/action experts. The purple loss box is schematic. Equation (22) gives the weighted action-aware spatial and temporal losses; the drawing alone does not specify their coefficients. e-architecturee-attentione-objectivee-figure-losse-latency
What it supports. The architecture supports an inference-efficiency claim because VGGT, the auxiliary expert and alignment modules are discarded. It also separates predicted training futures from the executable action output used at deployment. Figure 1 reports nearly identical action-chunk latency to Fast-WAM, consistent with this structural change being confined to training.
Where the evidence stops. Figure 2 writes an unweighted geometry-loss sum, whereas Equation (22) includes separate coefficients. Their values are not given. This overview also leaves precise parameter sharing and gradient routing between experts unresolved.
2. Motivation
2.1 The problem and the proposed response
Appearance-oriented video prediction may miss the object–gripper–target relations needed for precise manipulation. The paper asks whether evolving geometry can improve policy representations during training while preserving a lightweight observation-to-action deployment graph. e-problem
2.2 What this reading follows
A manipulation policy can generate plausible visual futures yet still miss the alignment needed to place one cube on another. MECo-WAM addresses that gap by making geometry a training resource: a frozen visual geometry encoder supervises a third expert, while the deployed video/action experts briefly gain access to current-frame geometric context. That access disappears during training, and the entire geometry branch disappears at inference. The useful question is therefore whether geometric knowledge survives in the retained representations. Read the architecture and attention mask first, then the benchmark results, representation probe and component ablation. The evidence supports modest performance gains, with important uncertainty about their robustness and exact transfer pathway. e-probleme-architecturee-attentione-distillatione-robotwine-depthe-ablation
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | WAMs |
| Architecture | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Insufficient evidence to decide
WAM, geometric representation learning and efficient inference are supported. Separate video/action experts with mixed attention establish multiple experts, but not two autonomous planner/controller systems. Training predicts video and actions; deployment predicts actions without future imagination or inverse dynamics. Dual-system and Other mechanisms/Outside quadrants therefore depend on whether catalog axes describe training or deployment; the source alone does not settle that interpretation. e-architecturee-attentione-objective
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Separate future supervision from the information available to an action
Begin with the deployed conditional policy: current observation, current robot state and instruction determine an action chunk. Future frames are available in demonstrations, but will not exist when the robot must choose its next action. MECo-WAM therefore assigns them training roles. The video expert denoises future visual targets, and the geometry expert learns selected future relations from frozen VGGT features. The attention mask prevents future action queries from reading future video or geometry. Their temporary extra context is g_0, which is computed from the current RGB frame. This distinction is what makes the auxiliary supervision compatible with action inference without future imagination. The action chunk is the policy output; future predictions are learning targets, not a separate planner whose trajectory the robot follows. e-architecturee-attentione-objective
Figure 3. The temporary geometry connection reaches the current observation only. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Rows are query slots and columns are keys, as the Q/K labels indicate. A white cell is masked; a colored cell permits attention. Follow the future action rows: they can read the clean visual anchor f_0 and action slots, but their columns for future video are white. Hatched cells appear only in the g_0 column for future video/action queries. They are temporary access to geometry encoded from the current frame, not access to future geometry. The right grid removes geometry and future-video slots. Equations (9–10) describe stochastic read activation, with the implementation reducing its probability from one to zero during the first half of training. e-attentione-implementatione-distillation
What it supports. The mask and caption agree about the causal boundary: future visual or geometric targets do not become action-side context. Deployment retains current visual conditioning and action denoising. Removing the training read edges progressively gives the policy practice operating with the same geometry-free interface it will use at test time.
Where the evidence stops. Visibility specifies forward information access, not the complete backward gradient graph. Future geometry remains confined to its branch; the paper does not fully document the parameter sharing and gradient paths that make its losses improve retained video/action features.
5.2 Teach relative structure and its change, then emphasize relevant regions
The geometry objective does not simply copy each teacher feature vector. After projecting predicted geometry, MECo-WAM forms pairwise feature distances and normalizes the valid relation entries. Matching those relations avoids requiring the teacher and student to use the same absolute feature coordinates. A second loss compares normalized changes in relations between consecutive selected keyframes. Video tokens also receive relevance scores from their similarity to a pooled, temporally aligned action representation. Softmax scores mix with a uniform prior; pair weights combine the resulting token weights. Thus spatial and temporal errors contribute more strongly at regions associated with the action while some scene-wide supervision remains. These are relations in learned feature space, and the source does not establish a conversion from their values to physical distance. e-distillation
5.3 Ask what survives when the auxiliary branch disappears
Reader interpretation: the key hypothesis is about a durable representation change, so an extra geometry branch alone is an incomplete explanation. The reported isolated-expert variant barely improves task success, while temporary read access helps more. The frozen-backbone depth probe asks a complementary question: can a matched head decode geometry from the retained video/action features without auxiliary tokens? Its qualitative examples support that possibility, but do not close the causal argument. The mask confines future geometry to its branch, losses supervise predicted geometry, and action-aware weights involve video/action representations. Determining the exact transfer route therefore requires explicit parameter-sharing and gradient documentation. A useful reproduction should measure those routes alongside control outcomes, rather than treating either a sharper depth image or a small success gain as sufficient proof. e-attentione-distillatione-depthe-ablation
5.4 Training and inference
During training
Wan2.2-TI2V-5B supplies the video backbone. The action expert has 30 DiT blocks, 24 attention heads, 128-dimensional heads and hidden width 1024 (about 1B parameters); the 4D expert has width 512 (about 0.45B). VGGT-1B is frozen. Chunks contain 33 robot steps, horizon 32 and nine video frames at a 4:1 action/video ratio. e-implementation
Video/action objectives use conditional flow matching. Training uses 1000 timesteps, shift 5.0, AdamW at learning rate 0.0001, weight decay 0.01, cosine decay, bfloat16 and gradient clipping 1.0, on 64 NVIDIA H20 96GB GPUs. e-objectivee-implementation
During inference
Remove VGGT, the 4D expert and alignment modules. Current observation, instruction and state condition action denoising through retained video/action experts, with 10 denoising steps and CFG scale 1.0 on one RTX 5090 32GB. No future-video rollout is required. The number of actions executed before re-observation is unspecified. e-architecturee-attentione-implementation
5.5 Implementation flow
- Construct three denoising streams
The video expert keeps a clean first-frame VAE anchor and noises future video targets. The action stream keeps current proprioception clean and noises future actions. VGGT supplies a clean current geometry anchor and future targets whose input slots are noised. Selected predicted geometry keyframes receive relational supervision. e-architecture
- Restrict information flow
Experts have separate query/key/value projections in masked mixed attention. Current anchors are self-only. Future actions read current visual context and the action branch, excluding future video. Temporary reads reach only current geometry; future geometry stays within its branch. Read probability falls from one to zero during the first half of training. e-attentione-implementation
- Distill relations and their evolution
Projected predicted geometry is compared with frozen VGGT using normalized pairwise feature distances. Video–action cosine relevance, mixed with a uniform prior, weights token pairs. Spatial loss matches within-frame relations; temporal loss matches normalized relation changes between consecutive selected keyframes. These losses supervise geometry predictions, not a deployment-time depth decoder. e-distillation
6. Experiments & results
MECo-WAM adds a training-only geometry expert to a video-action model, then removes it for action inference. Frozen VGGT features supervise action-weighted spatial relations and temporal changes; temporary attention to current geometry decays away. Reported gains are modest in simulation and task-dependent on a real robot, with essentially unchanged action-chunk latency (e-architecture, e-attention, e-distillation, e-robotwin, e-real, e-latency).
6.1 Read the original evidence
Table 2. The aggregate RoboTwin gain is concentrated in the clean condition. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the final two rows, which compare Fast-WAM and MECo-WAM without embodied-policy pretraining. Read Clean and Rand. separately before the Average column: the mean can conceal an uneven benefit across conditions. The retained note explains that P.T. means embodied-policy pretraining, so a cross does not mean the method lacks a pretrained video backbone. The setup says randomization changes object poses, appearance, clutter, illumination and tabletop layouts. Other rows provide context, but their pretraining indicators differ. In particular, the LingBot-VA average is a useful reported benchmark comparison, not a controlled ablation of MECo-WAM's geometric objective. e-protocole-robotwine-implementation
What it supports. MECo-WAM improves average success from 91.83% to 92.62%, a gain of 0.79 percentage points. Clean success increases by 1.38 points, whereas randomized success increases by 0.20. Its reported average also exceeds pretrained LingBot-VA's 92.20%, although the rows do not share the same pretraining history.
Where the evidence stops. The PDF delegates simulation configuration to Fast-WAM and does not supply rollout counts, seeds or confidence intervals. The small randomized gain therefore cannot establish a reliable robustness improvement from this table alone.
Table 3. Real-robot gains differ between stacking and size sorting. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Treat the two gray task bands as separate experiments. Within each, locate the Avg column under each metric: SR and PR are success and progress percentages, CR is correction count, and CT is completion time in seconds. Upward arrows favor higher rates; downward arrows favor fewer corrections and shorter time. The Test 1, Test 2 and Test n columns show selected trial entries, so do not recompute the published averages from just those displayed columns. Compare Fast-WAM and MECo-WAM within a task before comparing tasks. The surrounding setup identifies an ARX-R5 arm and says the camera configuration and execution budget are matched. e-protocole-real
What it supports. Stacking remains tied at 60.0% success and 75.0% progress, while corrections fall from 1.67 to 0.83 and time from 27.06 to 25.71 seconds. Sorting success rises from 60.0% to 70.0%, with progress 75.0% to 80.0% and time 38.49 to 31.96 seconds. The physical evidence is task-dependent.
Where the evidence stops. The source does not define n, all averaging denominators, correction criteria or missing-value handling. A dash must not be read as zero. These two cube tasks provide a narrow basis for broader real-robot generalization.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RoboTwin 2.0 manipulation Clean/randomized bimanual simulation following the cited Fast-WAM configuration; both compared methods lack embodied-policy pretraining. | MECo-WAM: clean 93.26, randomized 91.98, average 92.62. Task success rate (%) | Fast-WAM: 91.88, 91.78, 91.83; average gain 0.79 percentage points. Pretrained LingBot-VA averages 92.20. Clean gain: 1.38 points; randomized gain: 0.20. No uncertainty is reported. e-protocole-robotwin |
| LIBERO manipulation Spatial, Object, Goal and Long suites; each described as 10 tasks with 500 expert demonstrations. Reported Fast-WAM evaluation configuration. | MECo-WAM: 98.8 / 100.0 / 98.2 / 95.8; reported average 98.2. Task success rate (%) | Fast-WAM: 98.2 / 100.0 / 97.0 / 95.2; average 97.6. Both lack embodied-policy pretraining; pretrained LingBot-VA averages 98.5. A 0.6-point reported average improvement; not the table's highest overall average. e-protocole-libero |
| RoboTwin action-chunk inference latency Figure 1 latency comparison; implementation specifies one RTX 5090 32GB for inference. | MECo-WAM 198.73. Latency (ms) | Fast-WAM 198.65. Essentially matched latency supports unchanged deployment cost; timing variability and detailed timing boundaries are absent. e-latencye-implementation |
| Stack Cubes ARX-R5 tabletop manipulation; identical camera setup and execution budget across methods. | 60.0 / 75.0 ; 0.83 ; 25.71. Success / progress (%) ; corrections (#) ; completion time (s) | Fast-WAM: 60.0 / 75.0 ; 1.67 ; 27.06. Tied success/progress with lower corrections/time; averaging denominators are not explicit. e-protocole-real |
| Sort Cubes by Size ARX-R5 tabletop manipulation; three cubes arranged in a size-ordered line. | 70.0 / 80.0 ; 1.00 ; 31.96. Success / progress (%) ; corrections (#) ; completion time (s) | Fast-WAM: 60.0 / 75.0 ; 1.33 ; 38.49. Ten percentage points higher success; trial count and uncertainty remain unspecified. e-protocole-real |
| RoboTwin component ablation Table 4; authors report identical training and inference settings. | Full model: 92.62 ; 0.013. Uniform-weight variant: 92.38 ; 0.017. Average success (%) ; action-column entries labeled MSE (×10) | Fast-WAM: 91.83 ; 0.032; isolated 4D expert: 91.87 ; 0.034; decayed-read variant: 92.14 ; 0.026. The isolated branch contributes little; read access and action-aware weighting improve outcomes. MSE entries retain the source's unresolved scale. e-ablation |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 4. A frozen-feature depth probe tests whether geometry remains decodable after co-training. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare each example horizontally, beginning with the raw scene at left. The next two columns visualize depth readouts from Fast-WAM and MECo-WAM features; the final column is pseudo-ground-truth generated by VDA. Black rectangles identify local regions the authors emphasize, including boundaries around manipulated objects and nearby structures. The caption specifies a matched DPT-style head using tokens from the final four layers. The probing paragraph further states that the video-action backbones are frozen and auxiliary 4D tokens are excluded. That setup matters: a useful probe must read the representation that survives deployment, rather than obtain geometry directly from the discarded teacher or expert. e-depth
What it supports. The authors interpret the sharper structures in the MECo-WAM column as evidence that geometry is retained in deployed features. The probe complements task success because it asks a representation question rather than merely repeating a manipulation score. It is qualitative evidence for accessible geometric information.
Where the evidence stops. These colors have no numeric depth scale, and the reference is pseudo-ground-truth. No depth-error table, probe-training protocol or statistical uncertainty is supplied. Selected examples cannot establish metric depth accuracy or causally explain the control improvements.
Table 4. An isolated geometry branch gives little benefit; interaction and weighting matter in the reported ablation. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the rightmost success column first, then use the action and video columns as diagnostics rather than interchangeable outcomes. The isolated 4D expert barely changes success; the decayed-read row improves it more. Rows labeled with spatial and temporal losses examine geometric supervision, while Full w/o aware replaces action-aware weighting with uniform weights. The final row combines the proposed components. The video columns have different preferred directions: PSNR and SSIM rise, LPIPS falls. Crucially, preserve the action header exactly as MSE (×10). The surrounding prose repeats the displayed entries as MSE without explaining this scale, so the crop supports comparisons between entries but not an unqualified raw-error conversion. e-ablatione-attentione-distillation
What it supports. Reported success progresses from 91.83% for Fast-WAM to 91.87% with an isolated expert and 92.14% with decayed reads. Uniform weighting reaches 92.38%, below the full model's 92.62%. The full row also reports PSNR 30.72 and the smallest printed action-error entry, 0.013, versus baseline 0.032.
Where the evidence stops. The table lacks a complete factorial configuration matrix, uncertainty estimates and a clarified MSE scaling convention. Its ordering is consistent with the proposed mechanism, but does not isolate every interaction or prove that improved geometric representation causes the success gain.
7. Analysis & limitations
7.1 What the evidence leaves open
Small simulation gains lack reported seeds, confidence intervals and rollout counts. Real-world evaluation covers two cube tasks; Table 3 shows Test 1, Test 2 and Test n without defining n, missing-value handling or all metric denominators. e-protocole-robotwine-reale-ablation
Matched DPT-style depth probes use frozen final-four-layer features and VDA pseudo-ground-truth. Qualitative examples support decodable geometry, but provide neither quantitative depth accuracy nor proof that representation changes cause the success gain. e-depth
The mask isolates future geometry, while geometry losses and action-aware weights involve different expert features. Exact parameter sharing and gradient routing are insufficiently specified to establish how each loss changes deployed features. e-attentione-distillation
7.2 Questions for discussion
- Which parameter-sharing or gradient path transfers future-geometry supervision into deployed experts?
- Does action relevance outperform shuffled weights under matched geometry losses and training budgets?
- Do small randomized-setting gains persist across seeds and harder contact-rich tasks?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction requires the named backbones, paired trajectories, temporal sampling, masks and inference settings. Missing details include total training steps, batch size, loss coefficients, relevance temperature/uniform-mixture values, exact relation normalization, keyframe/token selection, action encoding, software versions and detailed data splits. e-architecturee-distillatione-objectivee-implementatione-protocol
Reader-proposed checks compare absent, constant and decaying geometry reads, then learned, uniform and shuffled action weights under matched settings. Measure success, latency, depth-probe error and gradients into retained experts; these checks have not been run. e-attentione-distillatione-depthe-ablation
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Does decaying geometry access preserve gains after branch removal?
Reader-proposed, not run: train matched variants with no current-geometry reads, constant read probability one, and the reported one-to-zero schedule. Keep the 4D expert, losses, initialization, optimizer budget and data identical. Evaluate every variant with VGGT, geometry expert and alignment modules removed, using identical RoboTwin seeds, rollouts and denoising steps. Report clean/randomized success with uncertainty, action error with explicit scaling, and synchronized action-chunk latency. Also verify that changing future geometry tokens cannot change action outputs under the stated training mask when current inputs are fixed. If decaying reads do not outperform absent reads after removal, or remain dependent on geometry at test time, the proposed transfer benefit is weakened. e-attentione-implementatione-robotwine-ablation
Check 2: Does action alignment help beyond generic geometric regularization?
Reader-proposed, not run: compare learned relevance, uniform weights and relevance computed from temporally shuffled action representations. Keep action-training labels unchanged, preserve total geometry-loss weight, and match keyframes, read schedule, training budget and evaluation rollouts. Measure RoboTwin success, the same frozen-backbone DPT probe on a fixed held-out set with VDA pseudo-targets, and geometry-loss gradient norms in retained video/action parameters, with and without detaching relevance weights. Report teacher-relative probe error separately from control. If shuffled relevance matches correctly aligned relevance within uncertainty, the claimed importance of action alignment is weakened; if gradient changes occur only through the weighting path, that narrows the explanation of transfer. e-attentione-distillatione-depthe-ablation
8.3 Reading coverage
Visual audit: All nine supplied PDF pages were rendered and visually inspected after all four text chunks were read. Page 1 establishes title/authors/affiliations/version and Figure 1 latency annotations; page 2 provides motivation and positioning; pages 3–5 establish architecture, mask, losses, training hardware and inference settings; pages 5–7 establish protocols, all four result tables and all representation/robot examples. Pages 8–9 contain references, not an appendix. Six final original crops were individually viewed: Figures 2–4 and Tables 2–4. Figure 2's simplified loss box was checked against Equation (22); Figure 3's row/column orientation, masked cells and current-geometry reads agree with its caption and equations. Table 4's MSE scaling remains unresolved and is disclosed. Table 2 retains its short note defining pretraining. No separate supplement, code or external configuration was inspected.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- PDF p. 1: title, authors, affiliations, arXiv version stamp, Abstract
- PDF pp. 1–2: Introduction
- PDF p. 2: Related Work, World Action Models and Geometry-Aware Embodied Models
- PDF pp. 3–5: Methodology, Problem Formulation, Multi-Expert Co-Training Architecture, Decayed 4D Read-Mask Attention, Action-Aware Temporal Geometric Distillation, Training Objective and Inference
- PDF pp. 5–7: Experiments, Experimental Setup, Main Results, Representation probing, Ablation Studies
- PDF p. 7: Conclusion
- PDF pp. 8–9: References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- All four supplied chunks and all nine PDF pages were read; visual inspection resolved the text extraction limitation for figures and tables in the retained PDF.
- Observed edition: arXiv:2607.05468v1 [cs.RO], 6 July 2026. Title and all seven authors match the catalog after name-order normalization. No other revision was supplied or compared; no edition difference is established.
- No appendix or dedicated limitations section is present in this PDF. No separate supplement was supplied.
- The printed project-page address was not opened. Code, checkpoints, external benchmark configurations and cited papers were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title block and left-margin arXiv stamp
Observed title: Learning 4D Geometric Priors for Inference-Efficient World Action Models. Authors: Jianjun Zhang, Jian Zhu, Taiyi Su, Chong Ma, Zitai Huang, Yi Xu, Hanli Wang. Affiliations: Tongji University; AIRC, Midea Group. Stamp: arXiv:2607.05468v1 [cs.RO], 6 Jul 2026.
Go to primary source ↓e-problemPDF pp. 1–2, Abstract and Introduction
The paper motivates action-relevant temporal geometry beyond appearance-oriented video latents while retaining the base model deployment graph.
Go to primary source ↓e-architecturePDF p. 3, Figure 2 and caption; Problem Formulation; Expert tokens; Eqs. (1–5)
Observation/state/instruction condition action chunks. Clean current anchors accompany noisy future video/action/geometry slots. Frozen VGGT supplies geometry targets; selected predicted keyframes receive loss. Geometry expert, teacher and alignment are removed at deployment.
Go to primary source ↓e-attentionPDF p. 4, Figure 3 and caption; Eqs. (6–10); Decayed 4D Read-Mask Attention
Expert-specific Q/K/V tensors enter mixed attention. Diagram and prose agree: anchors are self-only, future actions cannot read future video, temporary reads reach current geometry only, future 4D slots remain internal. Bernoulli activation decays linearly. Inference removes 4D and future-video slots.
Go to primary source ↓e-distillationPDF pp. 4–5, Action-Aware Temporal Geometric Distillation, Eqs. (11–19)
Projected geometry predictions match normalized teacher pairwise feature distances. Video/action cosine relevance is softmax-normalized, mixed with a uniform prior and combined geometrically for pair weights. Spatial and temporal relation discrepancies supervise geometry predictions. Numeric relevance settings and exact normalization are not given.
Go to primary source ↓e-objectivePDF p. 5, Training Objective and Inference, Eqs. (20–23)
Video/action flow matching interpolates clean targets and Gaussian noise and predicts epsilon minus target. Weighted spatial/temporal geometry loss joins weighted video/action losses; numeric weights are not stated.
Go to primary source ↓e-figure-lossPDF p. 3, Figure 2 loss box; PDF p. 5, Eq. (22)
Figure 2 writes L_4d as L_geo plus L_tem, omitting action superscripts and coefficients. Equation (22) explicitly includes alpha_geo and alpha_tem and action-weighted losses. The diagram does not establish unit coefficients.
Go to primary source ↓e-implementationPDF p. 5, Experimental Setup, Implementation details
Video: Wan2.2-TI2V-5B. Action: 30 blocks, 24 heads of dimension 128, hidden width 1024/about 1B parameters. Geometry: width 512/about 0.45B; frozen VGGT-1B. Read probability 1 to 0 in first half of training. Chunks: 33 steps, H=32, 9 video frames. Flow timesteps 1000, shift 5.0; AdamW 0.0001, weight decay 0.01, cosine schedule, bfloat16, clipping 1.0. Training: 64 H20 96GB; inference: one RTX 5090 32GB, 10 denoising steps, CFG 1.0. Robot: ARX-R5.
Go to primary source ↓e-protocolPDF pp. 5–6, Experimental Setup, Benchmarks
Simulation follows cited Fast-WAM configuration without restating the full protocol. LIBERO: four suites, each described as 10 tasks with 500 demonstrations. RoboTwin randomization varies poses, appearance, clutter, lighting and layouts. Real-world stacking/sorting use identical cameras and execution budget; numeric rollout counts and budget are not provided.
Go to primary source ↓e-liberoPDF p. 6, Table 1, Fast-WAM / MECo-WAM / LingBot-VA rows and caption
MECo-WAM success percentages: 98.8, 100.0, 98.2, 95.8, average 98.2. Fast-WAM: 98.2, 100.0, 97.0, 95.2, average 97.6. Both lack embodied-policy pretraining. Pretrained LingBot-VA average: 98.5.
Go to primary source ↓e-robotwinPDF p. 6, Table 2, Clean / Rand. / Average columns; RoboTwin 2.0 discussion
MECo-WAM 93.26/91.98/92.62 versus Fast-WAM 91.88/91.78/91.83; both lack embodied-policy pretraining. Pretrained LingBot-VA scores 92.90/91.50/92.20. No uncertainty is shown.
Go to primary source ↓e-latencyPDF p. 1, Figure 1, MECo-WAM and Fast-WAM annotations and caption
Action-chunk latency/success annotations: MECo-WAM (198.73 ms, 92.6%), Fast-WAM (198.65 ms, 91.8%). Axes are broken; exact labels are used rather than visual interpolation. Timing variation is absent.
Go to primary source ↓e-realPDF p. 7, Table 3, task-specific Avg columns and caption; PDF p. 6, Real-world evaluation
Stacking SR/PR/CR/CT: MECo-WAM 60.0/75.0/0.83/25.71, Fast-WAM 60.0/75.0/1.67/27.06. Sorting: 70.0/80.0/1.00/31.96 versus 60.0/75.0/1.33/38.49. SR/PR percentages, CR counts, CT seconds. Test n and dashes are not fully explained.
Go to primary source ↓e-ablationPDF p. 7, Table 4, all rows/column headers; PDF pp. 6–7, Ablation Studies
Avg SR / printed MSE entries: baseline 91.83/0.032; 4D expert 91.87/0.034; decayed read 92.14/0.026; spatial loss 92.22/0.022; temporal loss 92.25/0.019; Full w/o aware 92.38/0.017; full 92.62/0.013. Full PSNR/SSIM/LPIPS: 30.72/0.942/0.037. Header: MSE (×10); prose repeats entries without explaining scaling. No uncertainty or full factorial configuration matrix is supplied.
Go to primary source ↓e-depthPDF p. 5, Figure 4 and caption; PDF p. 6, Representation probing
Matched DPT-style heads use the final four layers of frozen video-action backbones. Columns: raw image, Fast-WAM, MECo-WAM, VDA pseudo-ground-truth. Probes exclude auxiliary 4D tokens. Sharper geometry is a qualitative author interpretation; no depth-error metric or color scale is supplied.
Go to primary source ↓e-graspPDF p. 7, Figures 5–6 and captions
Representative sequences show stacking and size sorting. Grasp images compare pi_0, Fast-WAM and MECo-WAM. These are illustrative examples, not a quantitative pose-sensitivity benchmark.
Go to primary source ↓8.5 Primary sources
Learning 4D Geometric Priors for Inference-Efficient World Action Models ↗
PDF · 5,067 extracted words
Source fingerprint
9049c3213a97ff288a58e07c482484fd2703a5cf103ec918b14cca5b1c284461